Periodic Labs：训练 AI 科学家，与 Liam Fedus 和 Ekin Dogus Cubuk 对谈（来自 a16z）

Today I'm pleased to share a special crosspost from the a16z podcast featuring a16z general partner Anjney Midha, who also recently joined me on The Cognitive Revolution to discuss sovereign AI. Today, we're in conversation with Liam Fedus, former vice president of post-training research and co-creator of ChatGPT at OpenAI, and Ekin Dogus Cubuk, former head of materials science and chemistry research at Google DeepMind. Together, they've co-founded Periodic Labs and just announced a $300 million seed investment led by Andreessen Horowitz.

Before diving in, a quick note: while Turpentine was recently acquired by a16z, my editorial independence remains unchanged. I'm sharing this episode simply because I think it offers a really valuable perspective on the future of AI-powered science. Regular listeners will no doubt notice some overlap between this conversation and our recent episode with Radical AI. Both companies believe that there simply isn't enough high-quality experimental data in the existing scientific literature to train foundation models for physics and chemistry.

Both have raised serious capital to build automated physical laboratories meant to connect AI-generated hypotheses directly to real-world experiments, using feedback from physical reality as the reinforcement learning signal. The goal is to teach AI models a form of scientific intuition and thereby accelerate scientific progress itself. Of course, there are still many possible ways to focus such an ambitious project. While Radical AI has recently announced a contract with the U.S. Air Force to develop high-entropy alloys for use in hypersonic aviation, Periodic Labs has set the goal of discovering a high-temperature superconductor as its North Star.

The expectation is that, to get there, they'll need to achieve countless subgoals along the way, including autonomous synthesis and autonomous characterization. Importantly, while the science and macro strategies are similar, the conversations are actually quite different. Whereas I tend to explore the technical details in arguably tedious depth, Anjney focuses much more on the human and organizational dimensions of building an AI-for-science company. As you'll hear, because no human comes close to holding all of the scientific knowledge and intuition that Periodic Labs hopes to train into its AI systems, the company prioritizes people with intense curiosity and mission alignment, and it doesn't require advanced degrees.

They take pride in their “no stupid questions” culture, and they host weekly teaching sessions in which ML researchers, physicists, and chemists can all learn from one another. Recognizing that even $300 million won't be enough to achieve their ultimate goals, and that even a wildly successful company is only one part of the broader scientific ecosystem, they have thoughtful plans to commercialize their progress in the form of an intelligence layer for advanced manufacturing companies while also starting a grant program, even at this early stage, meant to elicit key contributions from academia. Overall, I love the vision and ambition on display here, and I admire the conviction with which a16z and others are backing it.

While this doesn't come up in the episode, I've long believed that long-term AI safety might best be achieved by creating domain-specific superintelligences. That would mean that the AIs that advance fundamental science don't need to have an advanced theory of mind or persuasion skills. In any case, as much fun as I'm having playing around with Sora 2, it does seem quite clear that a future of truly radical abundance requires AI systems that go beyond the digital world and iterate directly against nature's own ground truth.

This is a conversation about building an AI research company meant to develop systems that autonomously explore and deeply understand the physical world.

Ultimately, science is driven by experiment in the real world. That's what we're doing with Periodic Labs. We're taking these precursor technologies and saying, “Okay, if you care about advancing science, we need to have experiments in the loop.”

The applications of building an AI physicist, for lack of a better word, that can design in the real world are so broad. You can apply it to advanced manufacturing. You can apply the materials science to chemistry—to any process where R&D with the physical world is required. It seems like we'll benefit from breakthroughs that Periodic Labs is working on. For example, if we could find a 200-kelvin superconductor, even before we make any product with it, being able to see such quantum effects at such high temperatures would be such an update to people's view of how they see the universe.

So, Liam, you were the co-creator of ChatGPT, and you were running some of the physics teams at DeepMind. Let's talk about how you guys met and what was the moment when you realized that you had to leave both of those labs to start Periodic.

I believe we met 8 years ago at Google Brain, flipping over a large tire.

Yep.

At Google. You've got to give us more on that story.

So, there was a gym at Google's facilities, and I think that's where Dogus and I met. There was this massive tire that a single person basically couldn't flip by themselves. Dogus was trying to flip it, and he pulled me over. He was like, “I think the two of us could do it.”

And why were you trying to flip this tire?

You know, why not?

But yeah, I tried doing it. I couldn't do it. Then I was like, “Who's the strongest person I can find?” It was either Barrett or Liam, and I picked Liam. It worked; we just flipped it.

And was that the moment when you guys both realized you had physics backgrounds? How did that happen? How did you go from flipping tires to flipping experiments?

Yeah. I don't know if Liam remembers this, but we would catch up over the years, and we would often end up talking about either quantum mechanics or superconductivity. This was very common, but I never thought we would end up working on physics together. Liam was working on LLMs, and they were going really well. I wasn't using LLMs, but I was noticing that LLMs were becoming more and more impactful in my work.

One way they were becoming impactful was when I was trying to remember things about chemistry and physics: I could just talk to the chatbot and actually learn a lot of stuff I'd forgotten. Another way was, of course, coding. We were writing simulations, and the LLM was so helpful in writing those simulations for us. So then the question was: can we use LLMs more as a first-class citizen in physics research?

Yeah. And I think, leading up to this decision to leave, Dogus and I were just connecting and talking about these different tech trees. We were looking at improvements in language models and reasoning. We were seeing what high-compute reinforcement learning could do. On the materials science side, we were seeing scaling laws within physics and chemistry, both with respect to simulations and with respect to experiments. It was the same kind of principles at play in ML.

To both of us, and to a lot of people in the field, the goal of this technology is to accelerate science and accelerate physical R&D. Chatbots were a great milestone along the way, but we really want to see technology out in the world.

And we felt like this was just the right place to begin. Physics is very verifiable. It's a great reward function with a fairly fast iteration loop. You have simulators for large classes of physical systems, and we felt like, in order to create this AI scientist, this is the beginning of that path. So we built that conviction and decided to found Periodic.

Well, let's take a second to talk about what Periodic is and what it does.

Periodic Labs is a frontier AI research lab that's trying to use LLMs to advance physics and chemistry. We feel like having experiments in the loop, tightly coupled with simulations and LLMs, is extremely important. So we're building up a lab that will generate high-throughput, high-quality data, and we will use LLMs and simulations in conjunction with experiments to try to iterate. Science, by its nature, is an iterative process, and we feel like LLMs, using all these tools that are available to humans, can do a great job accelerating physical R&D.

I'd say the objective is to replace the reward function from math graders and code graders that we're using today. For example, with math graders, you have a prompt: what is 2 plus 2? The ground truth is 4. You can put a lot of optimization pressure on problems like that that are programmatically checkable. What we're doing by having the lab is creating a physically grounded reward function that becomes the basis against which we're optimizing.

If a simulator has deficiencies or issues, we always error-correct, because for us, the ground truth is the experiment. The RL environment—nature—is our RL environment in this setting.

Let's take a second for folks who might not be familiar with what you mean by a lab that will verify RL in the real world. Can you talk a little bit about how experiments work? How are AI models trained today, and how is that different from how they're going to be trained, developed, post-trained, and deployed at Periodic? It might also be helpful to talk about how you created ChatGPT.

So, ChatGPT originally—the technology evolved very rapidly over the last few years.

When we were first creating it, it was a very standard RLHF pipeline. You have a pretrained model, and it’s sort of like a raw substrate. What you’re trying to do is take this auto-completion model and turn it into something useful.

The way we did it at that point was with supervised data. Given some input, we would say, “This is a desired output.” If we’re trying to get it to act as an assistant, we create some tuples like that. Then you run reinforcement learning, but now you’re learning against a reward function that’s trained against human preferences. Humans will say, “Given this input, I would prefer completion A to completion B.” You do that over and over again, and you can create a reward function that can then be optimized against.

That is sort of the basis of how we created ChatGPT. But then there’s a huge gap between the original model and what we have today. I think part of that is reasoning, but also part of that is just much better, more precise reward functions.

So, the reward functions that we were using originally couldn’t determine whether you were mathematically correct or not?

Early versions of ChatGPT were mathematically not particularly strong, and it sort of resulted from the reward function.

What did you optimize against?

The reward function basically encoded, “Be a friendly assistant. Try to help people get to their thing.” But it had no sense of whether this was mathematically correct or whether the code was valid or not.

We made huge advances in the correctness of our reward functions, but this is all digital. We’re creating tasks based on the internet, textbooks, and papers, and this is great—it lays a foundation. But ultimately, science is driven by experiments in the real world.

That’s what we’re doing with Periodic Labs. We’re taking these precursor technologies and saying, “Okay, if you care about advancing science, we need to have an experiment in the loop,” and that becomes our reward function for our agents. As Ekin was saying, our agents are doing the same types of things you would use for coding or to help answer a query, but now instead of just giving them tools like, “Here’s Python, here’s a browser,” we have tools like quantum mechanics to simulate different systems.

Ultimately, we’re going to a lab, and then that becomes the basis of what the system is optimizing against.

That’s sort of just the natural end state of these systems. People in AI often say “lab,” but what they’re referring to is quite different from what you guys mean by a lab. What’s the difference?

That’s right. As Ekin was saying, so far the LLMs have gotten really good at logic and math. There are verifiable rewards. What is the next frontier in terms of inquiry after logic and math? I’d say it’s physics.

When you say physics, there are different energy scales. There’s astrophysics, studying galaxies; there’s fusion and nuclear physics; but then there’s the energy scale of physics that’s more relevant to our lives, and that’s quantum mechanics—Schrödinger’s equation. This is where biology happens, chemistry around us happens, and materials happen.

We felt like our first lab should basically be probing that quantum mechanical energy scale. For us, that would be physics at the level of solid-state physics, materials science, and chemistry.

One of the more fundamental ways of making things around us is powder synthesis. You take powders of existing materials, mix them, and heat them up to a certain temperature, and it becomes a new material. That’s one of our labs. We’re going to have a powder synthesis lab, and it turns out this is one of those methods where robots can do it using very cheap, simple methods.

I don’t know if you saw this coffee-making robot at the SFO airport. A robot that’s basically at that level can mix powders and put them in a furnace. That’s a very rich field. Using that method, you can actually discover new superconductors, magnets, and all kinds of materials that are very important for the technologies around us.

But at the core of it, it’s just quantum mechanics. We feel like teaching these LLMs to be foundation models for quantum mechanics will be the next frontier for LLMs.

Why haven’t the models that are currently out in the world and deployed been able to do this?

Great question. As mentioned earlier, science is by its nature iterative. Even the smartest humans tried many times before they discovered the things they discovered.

Maybe this is one of the confusing points about LLMs: an LLM can be very smart, but if it’s not iterating on science, it won’t discover science. To be honest, humans won’t either. You put a human in a room without any chance to iterate on something, and they won’t discover anything important.

We feel like the important thing to teach these LLMs is the method of scientific inquiry. You do simulations, you do theoretical calculations, you do experiments, you get results, and the results are probably incorrect or not what you want at first, but you iterate on them.

We feel like that hasn’t been done yet. This is what we want to do, but we feel like you have to do it with the real physics, not just the simulation. This is why we have our own lab, where the LLM will have the opportunity to iterate on its understanding of quantum mechanics.

Fundamentally, machine-learning models are good at what you train them to do, and that’s sort of the nature of it. If a model is acting badly, you’re like, “Well, did you train it to do that task?”

Building on Ekin’s point, there’s sort of an epistemic uncertainty, this reducible uncertainty that you aren’t really building or collapsing unless you’re actually running an experiment. For instance, one of the engineers on our team was looking at a reported physical property in the literature, and it spanned many orders of magnitude.

If I train a system on that, these systems aren’t magic. The best they can do is replicate that distribution, but it’s really no closer to a deeper understanding of the universe—physics and chemistry.

Another point is that it’s very uncommon to publish negative results. All of the results are basically positive, and a valid negative result is very valuable. A negative result could be discarded because it was sloppy science, but there are valid negative results, and that’s a learning signal. This is something that our lab will produce as well.

So I think these three things—noisy data, no negative results, and the need for the ability to act in order to actually do science, which is an iterative endeavor—are the core theses of why we need a lab.

What might be the core way to measure Periodic’s progress against that goal, in your guys’ minds?

One simple one is high-temperature superconductivity. What is the highest-temperature superconductor we’ve synthesized? Today, the best number for ambient pressure is 135 Kelvin or so. We’ll know very easily if we’re doing well if we can go beyond that number.

That’s pretty fundamental. On the more applied side, there’s the processing of materials and its effect on the materials’ properties. We can just measure these properties directly. Let’s say it’s the ductility, toughness, or strength of the material.

As we measure it, the LLM will get a very clear signal. It’s hard to hack, unlike these other LLM training techniques. Really, what you see in real life is the signal that’s going to the LLM.

Yeah, effectively. Can you design the world around you? You need something with this property. Can this system discover and produce that, both from a fundamental scientific-discovery perspective and for an industry?

Someone’s working in space, defense, or semiconductors, and they’re having these issues. They’re trying to achieve this property of a material or this layer. Can the system accelerate the development of those technologies? It’s very grounded. That’s how we’ll know it’s working.

It feels like the applications of building an AI physicist, for lack of a better word, that can design the real world are so broad. You can apply it to advanced manufacturing, materials science, chemistry—to anything that involves a process where R&D with the physical world is required. It seems like it will benefit from breakthroughs that Periodic is working on.

Why hasn’t it been done before? And what is it about this moment in history that makes it the right time to attack this problem?

Maybe one component is difficult.

What makes it so difficult?

I think part of it is the team. In our view, this has been enabled by frontier technology in the last couple of years.

Ekin and I have been focused on putting together this N-of-1 team—this group of physicists, chemists, simulation experts, and some of the best machine-learning researchers in the world who have never been part of one concerted effort. We feel that, in order to actually achieve this, you need all this expertise. You need these pillars to do this.

So when you guys went about designing the team, after you left OpenAI and DeepMind, what was the primary heuristic that you used to guide yourselves in figuring out who you wanted on the team?

In terms of expertise, we wanted to have LLM expertise covered, experimental expertise, and simulation. For each of these, we wanted to have basically world-class talent.

Of course, for each team, there's actually a lot of subteams. It's like a fractal, right? Expertise is very fractal. For the experimental side, we wanted to cover solid-state chemistry, solid-state physics, automation, and the more facilities—the more operational aspects of experiments. On the simulation side, there's the more theoretical physics parts and the more coding aspects of simulations. And on the LLM side, of course, there's mid-training, RL, and infrastructure. For each of these, we tried to get basically the best people who have innovated in these subpillars.

The technology that we think is necessary to do this has really just emerged in the last couple of years. And this data isn't on a Reddit forum or something. You need to actually go produce experimental data and simulation data. It's siloed across all of these advanced industries, and many of them, while there's a desire, may not have knowledge of some of the most recent techniques that have been driving this recent wave in AI.

There was a moment in time when papers like the GPT-3 paper—for example, “Language Models are Few-Shot Learners”—proposed the idea of scaling laws. And then there was a follow-up paper, if you guys remember, from OpenAI called, I think, “Scaling Laws for Autoregressive Generative Modeling.”

That showed that, as long as you kept scaling up the amount of compute and data in the right combination, you could very predictably improve the performance of these models. The theory was that if you just kept doing that ad infinitum, there would be a bunch of emerging capabilities. These models would be able to reason about all kinds of problems out of domain and out of distribution.

Wouldn't that argue—how would you square the circle with that school of thought—that, naively, the current pretraining and post-training pipelines at most of the frontier labs will eventually crack physics as well? Why is this idea of physical verification so necessary, and is that school of reasoning wrong?

Excellent question. Scaling laws empirically seem to continue to hold, so that's not in question. But I think there's a question of what the y-axis is.

And that test distribution is very different from what we're talking about.

That test distribution—let's say you're pretraining on the internet—might be a representative sample from the internet, and you will have these predictable scaling properties. But that's not going to capture that you have a very different set of scaling properties with respect to different distributions.

Let's try to make this a little more concrete. Let's say, hypothetically, we're training a coding model and we have unit tests to provide some reward signal. The model writes some PR. We check that the unit tests go from failing to passing, and we say this was successful. We're going to reinforce these things.

You might say you start optimizing this, and now the system is becoming ever more capable of writing code for its own development, and you have this acceleration. You have this kind of takeoff scenario. Code is one of the most promising areas for this because there's an abundance of data online. You have this feedback loop where the system itself can begin to improve itself. It's a very promising technique, and we're all seeing the benefits of advanced coding models. It's accelerating quickly.

However, that model is not going to then cure cancer. The knowledge simply doesn't exist. You need to optimize against the distribution you care about. So that model, while it's going to be a very valuable tool as a software engineer, may help a cancer researcher do their analysis, but it simply doesn't have the data, the knowledge, or the expertise to iterate against that environment.

And I think that's just sort of the fundamental belief we have. Yeah, I mean, Liam and I worked on this a bit when we were looking at scaling laws for vision models, and this also came up a lot in the CLIP paper from OpenAI. The in-domain generalization and the out-of-domain generalization are monotonically correlated, but it's not necessarily linear.

What that means is you can keep improving your model, and it will improve as a power law in-domain. For out-of-domain tasks—by which I mean the things that you're trying to do that are a bit different from what's in your training set—it will also improve as a power law, but the slope of the power law may not be good enough. You might need to spend centuries before you get to the result you want.

We saw this in the NeurIPS paper, for example. We published a paper where we saw that, as you increase the size of your training set, the IID performance—the in-domain performance—improves as a power law. Out-of-domain performance also improves as a power law, but depending on what the out-of-domain is—how far you are from the training distribution—that power law might have such a small slope that it's basically useless.

This is one of the reasons we feel like the best way to make progress is to make your target as close to your in-domain training set as possible. The best way of doing this is to basically iterate on changing your training set to be more like what you want to do.

That's one answer. The other one is actually maybe even simpler: the experimental data we want actually doesn't exist. For example, if you want to learn on the experimental data in the literature for synthesis, it turns out the formation enthalpy labels—which is the energy it takes to basically assemble the atoms in the shape you want—are so noisy that, if you train a machine learning model on them, it's not predictive enough to predict the next one.

One of the reasons for this is, as Liam mentioned, people don't usually publish negative results, and negative results are usually very context-dependent. So what's a negative result for someone might be positive if they do things differently.

Not only is there this domain-shift problem, where what you're trying to do might be different from your training set and the power law won't have the large enough slope you want, but the other problem is that, for some of these things we want to do, there's no data. For example, for superconductivity, there's a lot of datasets you can look at, but the noise floor on them is so high that training on them usually doesn't help.

Liam, me, and the entire team are deep believers in scaling up and scaling laws, but just do a baseline for the thing you care about.

Is there a tension between being super bullish and just throwing more compute at the problem and the domain-specific pipelines that the lab you guys just described will have to focus on? In the case of Periodic, I think you mentioned the first baselines you guys are making are toward superconductivity and magnetism. What is it about those domains that make them good candidates for the first few pipelines that Periodic is working on?

Are they just pit stops along the way to an AI physicist that generalizes across all kinds of domains, or is there a danger of them being essentially off-ramps that don't result in the AI scientific superintelligence that is the north star for what you guys are doing?

Yeah. For example, I feel like the goal of high-temperature superconductivity is actually a goal that has so many subgoals in it. It’s a bit like when DeepMind and OpenAI started and said, “We’re going to do AGI,” but what they meant was that they had to do so many things before they got to these cool results.

For us, if you want to get a high-temperature superconductor, we probably need to get good at autonomous synthesis and autonomous characterization. We need to get good at characterizing different aspects of the material, using the LLM to run the simulations correctly. So it’s a north star, and there are so many goals along the way that would be, I think, impactful for the community.

That’s one reason. Another reason is that I feel like high-temperature superconductivity is such a fundamentally interesting question. For example, if we could find a 200-kelvin superconductor, even before we make any product with it, that in itself says so much about the universe that we didn’t know yet. To be able to see such quantum effects at such high temperatures, I think, would be such an update to people’s view of how they see the universe.

So we feel like it’ll be really impactful for humanity even before we make a product out of it. I think that’s one of the reasons. A technical reason also is that superconductivity is a phase transition, so it’s pretty robust to some of these details that we cannot simulate yet. For example, when you make the material, the superconducting temperature usually is more dominated by its crystal structure’s fundamental properties than by defects or microstructure, whereas there are certain other material properties where, even if the crystal has the property you want, there are so many other factors that you cannot simulate that would prevent you from seeing that properly.

So superconductivity has this nice philosophical upside and this technical upside to it. And it really rallies both the physicists—there are people who have studied physics for 40 years and are really excited about superconductivity—and there are people who never studied physics but are very excited about superconductivity. It’s quite rare to find a topic that unites the whole team.

Yeah. I mean, like Ekin said, in order to do this, there are so many foundational pieces to solve, and our tactic is that, in order to actually get to this goal of an AI scientist, you need to make contact and do the full loop somewhere. If you say you’re doing this in just very vague terms, you sort of just end up back on arXiv papers and textbooks.

So it’s really important for us to do the loop, but then create this repeatable process: How do you go from subdomain to subdomain? And there are really interesting questions about how well the ML systems generalize between these things. What is the generalization of a system between superconductivity data and magnetism data, for instance? And maybe that looks very different from its ability to generalize to fluid mechanics. I think there are fundamental arguments to make there.

But the goal is to create this repeatable system, prove it, and then just go through the different domains that way.

So, I can see the argument for why cracking room-temperature superconductivity from an experimental basis is extraordinarily valuable for humanity. But you guys are building a startup, and to use an analogy for why you need to have a clear medium-term path—or short- to medium-term path—along the way to a north star that is both commercially viable and net positive to society, what we’ve seen, for example, with other frontier labs that are working on automating white-collar work or software knowledge work is that there’s this north star of an AI researcher, but that along the way there were a bunch of subgoals and so on.

But a concrete kind of application that opened up a ton of commercial value and benefits for users on the way to that AI researcher was the idea of AI programming. Software engineering has become probably the first major domain, and that’s caused people to really update their priors about how useful AI models are beyond consumer applications. In terms of productivity, their impact has been extraordinary in just a few short months.

So if the traditional frontier labs’ north star was an AI researcher, and the path along the way to get there was programming—AI programming—what is that for Periodic? Basically, copilots for engineers and researchers in advanced industries.

Maybe just being in Silicon Valley, we really think about computer-oriented work. Everything is digital; everything is bits. But there are so many industries, as we were kind of talking about—a few of them are space, defense, and semiconductors—where they’re dealing with iteration of materials and physics, and that’s part of their workflow. How are they designing these new technologies, these new devices? In the absence of data, in the absence of good systems, they don’t really have particularly good tools.

That is our opportunity, and these are massive R&D budgets. So, yeah, while high-temperature superconductivity is a great north star, we very much understand that technology and capital are intertwined. We’re going to be able to maximally accelerate science if this is a wildly successful commercial entity.

To do so, we want to accelerate advanced manufacturing in all these different industries, become an intelligence layer for all these teams to accelerate their workflow, start reducing their iteration time, get them to better solutions more quickly, and accelerate their researchers and engineers.

Let’s dig a little bit deeper into that in practice: sort of a day in the life of a Periodic team member. Let’s say roughly half the team are ML scientists with machine-learning backgrounds, and the remaining half are physical scientists with physics or chemistry backgrounds.

How do you start by uniting the cultures? How do you take somebody whose primary career so far has been doing experiments in wet labs, doing physics and chemistry, and give them an intuition for ML, and vice versa? Because you guys are both physicists who then had the career trajectory where you also had the chance to be at frontier AI labs and were part of training systems that are now considered landmark machine-learning systems, like ChatGPT and GNoME.

For others who might be coming from one domain, how do you get the team to build an intuition for the other?

Yeah, so this is a great question, and we feel like it’s actually crucial for us to make sure these teams work very closely with each other. One of the things we’re seeing is that the physicists and chemists need to figure out how to teach the LLM how to reason about these things, because I think the frontier AI labs have figured out how to train them on math and logic, but not yet on physics and chemistry.

One thing we’re seeing that’s been really productive is that the physicists and chemists are thinking about what steps we should include in the mid-training and RL training that will teach the LLM how to reason correctly about quantum mechanics and how to reason correctly about these physical systems.

Another one, of course, is that the LLM researchers are learning quite a bit about the physics, the simulation tools, and the goals. So they’ve been working together really well. We have weekly teaching sessions where the LLM researchers teach how the RL loops work and how the data cleaning works, and then the physicists and chemists teach about different aspects of the science. The history of science is also very important.

So we feel like that’s been going really well. One way of looking at this is that the things we have to teach the LLM to be able to discover, say, a superconductor, include being able to read the literature really well—read all the papers and textbooks, find the relevant parts—and then being able to run simulations and theoretical calculations, and then take action: run experiments and learn from that.

We feel like this is quite similar to what physical R&D researchers in these companies have to do. They have to read the literature, read maybe internal documents or external documents, and then run simulations, run theoretical calculations, actually attempt the thing experimentally, and learn from that.

So we feel like all the progress we're making toward our internal superconductivity or physics goals is actually making our LLM better at serving our customers, who are doing very similar workflows.

Yeah, I think just culture: no stupid questions. You can ask just the dumbest physics question, the dumbest ML question. And I mean, there are a few faculty members as part of our company, and they are actually excellent teachers.

So these learning sessions have been really fantastic. Another thing I noticed is that computer scientists often think in terms of APIs. Scientists will say something, and they're always trying to map it. You're like, okay, what's the input? What's the output? What's the target? How do I map that back? It's always just this translation.

I think we have also built up, as part of the team, people on these different edges. So if you have a simplex of pure ML/LLM, pure experimentalist, and pure simulation, there are people who live inside this as well. They've been excellent bridges for translating between these different groups of people. It's active learning to learn the other spaces, creating APIs, and then these bridge-connector people. I think Dogus is an excellent example of that.

Is it a requirement for somebody who wants to join Periodic to have an advanced degree in physics or chemistry?

Absolutely not. One of the jokes we're making is about the NBA player who was saying, “I'm much closer to LeBron James than you are to me.” We were saying the opposite of that to candidates, because the amount that even our best physicist doesn't know about physics is much bigger than the amount that they do know about physics.

For a new candidate, even if they have no background in physics, how much they have to learn about what we're trying to do is actually not that different from how much the best physicist has to learn, because there's so much chemistry and so much materials science to learn. I think this is one of the interesting aspects of science today. In the past, in the 1800s, there were physicists who could do so many different things at the frontier.

Today, we've reached a point where our intellectual knowledge is so large that a leading thinker can usually only advance in one very specific field. Maybe this is actually holding us back, because to discover an amazing superconductor, as we keep going back to this example, you have to know so much about chemistry, physics, synthesis, and characterization. Unfortunately, I don't think any human knows enough about all of these, so we have to collaborate.

I think our team is a small example of this where, as Liam said, we have a lot of different points in that simplex. For any person, there's so much to learn, but that's true for basically every other scientist. For example, I supposedly come from the physics side of it, but I've been learning so much more physics because we now have people from different areas of chemistry and different areas of physics on the team.

I think it's true for LLM researchers as well. They come in, and there are aspects of LLMs that they probably didn't know until they started working with other researchers on our team. I think it's great, and it's a small example of what we're trying to do with the LLM, because we're trying to teach this LLM all these different things that we're learning as researchers.

It's like a really fun experience, I think. Yeah.

What are you finding makes a great researcher at Periodic that's different from what might make a great researcher at OpenAI, Anthropic, or DeepMind?

I would say there's very high overlap, but probably one of the biggest determinants is: do you care about this mission? Is accelerating science, to you, the big goal? Looking at the team right now, it's just an incredibly mission-driven set of folks who are like, yeah, this is the north star. Let's do that.

If someone really wants to improve some megacorp's products, you'd probably be better off at that megacorp, iterating and improving their products. But if you care about scientific discovery, I think Periodic Labs is the best place to do that.

How big is the team today?

We're roughly 30, I believe.

Yeah. And as you think about taking a lot of the research that's going on at the company and deploying that out in the real world, the kinds of customers that we've talked about—space, defense, and advanced manufacturing—these are mission-critical industries that are known for being essential to whatever part of the economy they're part of, but often they're not the fastest to adopt new technology.

How do you think about deploying the kinds of frontier agents that we've talked about, that are great at science and great at physics, in companies or organizations that might not be anywhere close to as sophisticated as you are in AI or ML? Do you have a working thesis for how to make sure that the arc of progress is not bottlenecked on deployment?

It sounds like you have a fairly good thesis on how to unblock the arc of scientific progress on the research side, but when it comes to deployment, what might be a working theory that you guys are optimistic about that would help get the systems that Periodic is building out into the real world?

Well, maybe one thing that we've noticed in our conversations with all these companies is that they are all looking for their AI strategy. They understand that the technology is shifting really quickly, and they're looking at how they're doing their work, and it's not changing as quickly as they think it should be.

Some industries are also losing key expertise in different fields. They're losing senior engineers and senior researchers, and they're asking, okay, how do we preserve that?

One thesis is understanding—thinking about these APIs, thinking about what the evaluations are, and thinking about what the biggest bottlenecks are for these companies. Looking at some of the problems they face, we can map that to our systems and say, well, we think we can dramatically accelerate this.

So it's not coming in and saying, hey, we're going to transform your fab line on day 1, or we're going to transform how you're doing everything. Forget everything. It's like, no, we're going to solve a really critical problem: well-scoped, with very clear evaluations. You co-draft that with them and just show them how powerful this technology can be when you optimize against the thing you care about.

So, nothing particularly surprising here, but a sort of land-and-expand type method, as you might expect.

But really looking for who the biggest promoters are within that company and what the biggest problems are. Make sure you're solving a very real thing for them, and intersect that with where our technical capability is highest.

You were on a call this morning with one of the customers in your pipeline. We don't need to name who, but what were some of the things you heard as their most urgent problems that they'd like Periodic to solve?

One of them was simulations. They spend a lot of time training people on some of these simulations they need to use, which are critical for their development. Being able to automate those simulations, I think, would be quite enabling for the design process.

Then there are some of the small things, like matching the formats and being able to feed the simulation results into the design pipeline. All of these seem quite important, and then being able to treat the data together in the same place. What else?

Well, I think there's a really fundamental question. A lot of these companies will rely on retrieval. That's sort of a super-lightweight thing. Someone shows up with a neural net, and they're like, great, we'll just retrieve over all of your data, and then that's your solution.

However, as we've seen with things like ChatGPT and other systems, it's when you pre-train on the data, when you actually encode the knowledge into the weights, that it's not just a retrieval system. You have a richer, deeper understanding of the material.

I think this is a big fundamental challenge. For instance, for this customer, they can give privileges to their employees and have retrieval act on their behalf—the system acts as the user—and so you can match those same kinds of privileges for access.

But if you start doing pre-training or mid-training on different parts, it's like, well, if you pre-train on every piece of data that might only be accessible to, say, the CEO of that company, then you have to figure out how to bucket that knowledge and create different types of systems.

Right now, after talking with the user, they don't seem to have a great solution for distilling all of the knowledge into a single model or into a set of models. So, going beyond retrieval to proper training. I think the supervised training they're doing is also really akin to the early days of ChatGPT, where it's input-output: you have a few examples, and you're kind of transforming this new way of thinking, which is, no, high-compute reinforcement learning is really effective.

This is how you should think about the strategies it's using. This is how you create effective tool use toward those problems, and this is how you optimize it effectively.

Could you describe for folks who may not be familiar with it what you mean by mid-training? People are familiar with pre-training, and they're familiar with post-training, but in the Periodic context, what does mid-training mean?

Yeah, sorry for the lingo.

So, I think this term came up years ago. It was like, well, we had pre-training, we had post-training, but sometimes you need to put in a little bit more knowledge.

Before search worked really well, there was an issue of freshness. We had pre-trained models, and they had a knowledge cutoff. There was a scrape of the internet at that point, but users wanted more real-time knowledge. So it was, how do you get that in there? Enter mid-training.

Mid-training is basically taking new data, new knowledge that’s not in the model, and continuing to pre-train. This differs from standard post-training, where post-training typically is more reinforcement learning and supervised learning. The mechanism—or the goal of it—is just to put a lot of knowledge into the model that didn’t exist before.

That’s mid-training in a nutshell. In the Periodic context, does that mean essentially going and injecting a ton of custom data from an experimental implementation in a particular customer or a particular industry? What are the lines—the atomic units—that you guys think of for mid-training that will improve the capabilities of the models on problems that they’re just terrible at today?

It’s all the knowledge. You can have very low-level descriptions of physical objects, like crystal structures, for instance. You can also have higher-level semantic descriptions of how a material was made, such as material XYZ. Trying to get all this data into the model is really valuable. It’s simulation data, experimental data—none of this exists. Basically, putting that knowledge into the model and making sure that these distributions are connected in some way is important.

What I mean by that is, if you just mix together distributions A, B, and C, there’s no guarantee of generalization. What you want to hope to see from these systems is that the inclusion of this other dataset is improving performance on the other datasets. These are just machine learning techniques or machine learning problems to solve. Basically, just make it an expert in physics and chemistry, where it was deficient before.

You guys both know that I spent some time running evals on a bunch of these models at the Stanford Physics Lab earlier this year, and the results were that the models are terrible at scientific analysis because they weren’t trained to do so.

On the other hand, many of the existing research teams working on the general models are investing in trying to make these better. Is there something about the way you’re building Periodic that lets you draft off all of that progress in the base models, or do you have to start everything from scratch and therefore not be able to be composable with advancements happening in the mainline models today?

Yeah, we benefit from a lot of different advances. One of them is that LLMs are getting better, and we definitely benefit from that because we take a pre-trained model and then mid-train it with high compute. Another one is that physical simulation tools are getting better. They’re open-sourcing new ways of simulating and new ways of using machine learning to predict properties. So we get to basically utilize all of those.

It seems like machine learning has made such an impact in the physics and chemistry fields that we expect these improvements to continue. I think another thing is that when we think about tools for agents, we think of, like, here’s a browser, here’s Python, but increasingly people think about tools as other neural nets, as other agents.

If you look at a lot of physics code, it’s not particularly deep. This isn’t competitive programming; these are kind of hacky scripts, but you can rely on some of the best systems for whatever they specialize in. So a neural net as a tool for these agents is something that immediately accelerates our work. You don’t have to replicate everything.

There’s a historical pattern that a lot of the fundamental research in the physical sciences we’re talking about here—physics, chemistry, and biology—has historically been done at university labs. Is there a role at all that you think the university ecosystem will play in Periodic’s future, or do you think these are just completely divergent paths?

Absolutely. So much of the simulation tooling we use has been developed in academia. Much of it is in Europe, for example, and a lot of the novel synthesis methods come from there, so we definitely benefit from a lot of this deep technical progress.

For example, a lot of physical simulation tools are complicated Fortran code that, in our team, we don’t really know how to develop very efficiently. But we feel like there’s definitely a very deep connection between academia and industry labs. Recently, a lot of the large-scale simulations have been done in industry labs like Microsoft, DeepMind, and Meta, but many of those tools were actually developed in academia and then passed on. So there’s actually a really nice synergy there.

I’d add a few other things, too. You found when you were evaluating models on their ability to do scientific analysis that they were deficient. This was probably not a direct goal for those teams training those models.

I think academia and these collaborations will help inform us: What are the important tasks? How do you do this analysis? What skills do we want to put in the model? A skill could be a full analysis, or it could be a smaller primitive as part of a larger analysis, but also, secondarily, it’s how you think.

One of the physicists was looking at the reasoning strategies of one of our models. He was like, “It’s all wrong. It’s all wrong.” And we were like, “What do you mean?” He was like, “No, this should be thinking at a higher level. It should be thinking in terms of symmetries.” This is the book that encodes the thinking strategies that will be more effective.

Of course, your reinforcement learning environment needs to reward those types of strategies. But given that some of the most premier scientists are using these strategies, they’re likely effective. These are the types of things where an industry-academic partnership can be so powerful, because industry is simply blind to these types of analyses, these tools, as well as this way of thinking.

Yeah. There’s also a way of connecting that to the tool in question, because language is very important, but in the human brain we also see a lot of visual, geometric processing. It’s plausible that while these LLMs will keep getting better and better, they’ll actually benefit from having geometric reasoning that’s separate.

Today, we can do that with equivariant graph neural networks. We can do it with diffusion models that are geometric tools by construction, and the LM can call them. Then it can have both the language aspect, which is very good for, say, a synthesis recipe, and the geometric aspects, which are very good for representing atoms and just designing geometries in general.

So how are you thinking about deepening Periodic’s ties with academic labs?

Yeah, this is very important for us. We have 2 major initiatives in this direction. One of them is that we’re starting an advisory board. This will include expertise spanning from superconductivity to solid-state chemistry to physics, and we want to make sure we’re in touch with these long-term research directions.

A lot of important government funding goes to these groups, and we want to have a tight coupling between what’s important for them and for us. This includes superconductivity expertise, such as Zhi-Xun Shen from Stanford on the experimental side and Steven Kivelson on the theory side. We also have synthesis expertise on the advisory board from Mercouri Kanatzidis from Northwestern University, and Chris Wolverton on the high-throughput DFT side.

Our second initiative is going to be through a grant program. We really want to enable some of this amazing work going on in academia, and some of that work isn’t a good fit for industry. It’s best done in academia. So we want to accept grant proposals and enable and support the kind of work that’s going to help the community, especially in relation to LLMs, agents in synthesis, materials discovery, and physics modeling.

So, for people who might be interested in joining Periodic, what are you guys looking for?

First off, someone deeply curious. Someone who really wants to understand machine learning and the science at a deeper level, who wants to make contact with reality, who wants to advance science—this has to be a driving thing.

But also pragmatic. What we’re trying to do is incredibly challenging, and we want someone who has a very careful process, is solution-oriented, and gets to goals quickly. And really, someone world-class along some dimension.

We’re looking across all these different pillars: machine learning, experimentalists, simulation, and people who can bring some sort of innovation to how you create a creative ML system. How do you bring new types of tools or new types of thinking to some of these state-of-the-art models? We’re looking for someone who can advance simulations and make them more robust and more reliable with experiments.

Yeah. And maybe one more thing I’d add is that Liam and I have been really looking for a sense of urgency in candidates, because we want these technologies not in 10 years.

You know, we don't want these LLMs to start improving science in 10 years, but we want them ASAP. So if the candidate feels a sense of urgency about improving these physical systems, discovering these amazing materials, and innovating on superconductivity, they would be a good fit.

Yeah. If you match all these, please reach out.

All right, sounds like we've got to amp up the speed and scale of what's happening at Periodic. We'll put the career links in the show notes. Thanks for coming, guys.